Advanced configuration options for ECI

Note

This feature is available with Docker Desktop version 4.27 (and later) on Mac, Linux, and Windows (Hyper-V). For Windows with WSL 2, this feature requires Docker Desktop 4.28 and later.

This page describes optional, advanced configurations for ECI, once ECI is enabled.

Docker socket mount permissions

By default, when ECI is enabled, Docker Desktop does not allow bind-mounting the Docker Engine socket into containers:

$ docker run -it --rm -v /var/run/docker.sock:/var/run/docker.sock docker:cli
docker: Error response from daemon: enhanced container isolation: docker socket mount denied for container with image "docker.io/library/docker"; image is not in the allowed list; if you wish to allow it, configure the docker socket image list in the Docker Desktop settings.

This prevents malicious containers from gaining access to the Docker Engine, as such access could allow them to perform supply chain attacks (e.g., build and push malicious images into the organization's repositories) or similar.

However, some legitimate use cases require containers to have access to the Docker Engine socket. For example, the popular Testcontainers framework sometimes bind-mounts the Docker Engine socket into containers to manage them or perform post-test cleanup. Similarly, some Buildpack frameworks, for example Paketo, require Docker socket bind-mounts into containers.

Starting with Docker Desktop 4.27, admins can optionally configure ECI to allow bind mounting the Docker Engine socket into containers, but in a controlled way.

This can be done via the Docker Socket mount permissions section in the admin-settings.json file. For example:

{
  "configurationFileVersion": 2,
  "enhancedContainerIsolation": {
    "locked": true,
    "value": true,
    "dockerSocketMount": {
      "imageList": {
        "images": [
          "docker.io/localstack/localstack:*",
          "docker.io/testcontainers/ryuk:*",
          "docker:cli"
        ]
      },
      "commandList": {
        "type": "deny",
        "commands": ["push"]
      }
    }
  }
}

As shown above, there are two configurations for bind-mounting the Docker socket into containers: the imageList and the commandList. These are described below.

Image list

The imageList is a list of container images that are allowed to bind-mount the Docker socket. By default the list is empty (i.e., no containers are allowed to bind-mount the Docker socket when ECI is enabled). However, an admin can add images to the list, using either of these formats:

The image name follows the standard convention, so it can point to any registry and repository.

In the example above, the image list was configured with three images:

"imageList": {
  "images": [
    "docker.io/localstack/localstack:*",
    "docker.io/testcontainers/ryuk:*",
    "docker:cli"
  ]
}

This means that containers that use either the docker.io/localstack/localstack or the docker.io/testcontainers/ryuk image (with any tag), or the docker:cli image, are allowed to bind-mount the Docker socket when ECI is enabled. Thus, the following works:

$ docker run -it -v /var/run/docker.sock:/var/run/docker.sock docker:cli sh
/ #

Tip

Be restrictive on the images you allow, as described in Recommendations below.

In general, it's easier to specify the image using the tag wildcard format (e.g., <image-name>:*) because then imageList doesn't need to be updated whenever a new version of the image is used. Alternatively, you can use an immutable tag (e.g., :latest), but it does not always work as well as the wildcard because, for example, Testcontainers uses specific versions of the image, not necessarily the latest one.

When ECI is enabled, Docker Desktop periodically downloads the image digests for the allowed images from the appropriate registry and stores them in memory. Then, when a container is started with a Docker socket bind-mount, Docker Desktop checks if the container's image digest matches one of the allowed digests. If so, the container is allowed to start, otherwise it's blocked.

Note that due to the digest comparison mentioned in the prior paragraph, it's not possible to bypass the Docker socket mount permissions by retagging a disallowed image to the name of an allowed one. In other words, if a user does:

$ docker image rm <allowed_image>
$ docker tag <disallowed_image> <allowed_image>
$ docker run -v /var/run/docker.sock:/var/run/docker.sock <allowed_image>

then the tag operation succeeds, but the docker run command fails because the image digest of the disallowed image won't match that of the allowed ones in the repository.

Docker Socket Mount Permissions for derived images

Note

This feature is available with Docker Desktop version 4.34 and later.

As described in the prior section, admins can configure the list of container images that are allowed to mount the Docker socket via the imageList.

This works for most scenarios, but not always, because it requires knowing upfront the name of the image(s) on which the Docker socket mounts should be allowed. Some container tools such as Paketo buildpacks, build ephemeral local images that require Docker socket bind mounts. Since the name of those ephemeral images is not known upfront, the imageList is not sufficient.

To overcome this, starting with Docker Desktop version 4.34, the Docker Socket mount permissions not only apply to the images listed in the imageList; they also apply to any local images derived (i.e., built from) an image in the imageList.

That is, if a local image called "myLocalImage" is built from "myBaseImage" (i.e., has a Dockerfile with a FROM myBaseImage), then if "myBaseImage" is in the imageList, both "myBaseImage" and "myLocalImage" are allowed to mount the Docker socket (i.e., ECI won't block the mount).

For example, to enable Paketo buildpacks to work with Docker Desktop and ECI, simply add the following image to the imageList:

"imageList": {
  "images": [
    "paketobuildpacks/builder:base",
  ],
  "allowDerivedImages": true
}

When the buildpack runs, it will create an ephemeral image derived from paketobuildpacks/builder:base and mount the Docker socket to it. ECI will allow this because it will notice that the ephemeral image is derived from an allowed image.

The behavior is enabled by default. It can be disabled by setting allowDerivedImages=false in the admin-settings.json file. In general it is not recommended that you disable this setting unless you know it won't be required.

A couple of caveats:

• The allowDerivedImages setting only applies to local-only images built from an allowed image. That is, the derived image must not be present in a remote repository (because if it were, you would just list it's name in the imageList).

• For derived image checking to work, the parent image (i.e., the image in the imageList) must be present locally (i.e., must have been explicitly pulled from a repository). This is usually not a problem as the tools that need this feature (e.g., Paketo buildpacks) will do the pre-pull of the parent image.

• The allowDerivedImages setting applies to all images in the imageList specified with an explicit tag (e.g., <name>:<tag>). It does not apply to images specified using the tag wildcard (e.g., <name>:*) described in the prior section, because Docker Desktop needs to know the tag in order to perform ancestor-descendant image checks. Therefore, if you want Docker socket mounts to be allowed for images derived from a parent image in the imageList, make sure the parent image is listed with name and tag.

Command list

In addition to the imageList described in the prior sections, ECI can further restrict the commands that a container can issue via a bind mounted Docker socket. This is done via the Docker socket mount permission commandList, and acts as a complementary security mechanism to the imageList (i.e., like a second line of defense).

For example, say the imageList is configured to allow image docker:cli to mount the Docker socket, and a container is started with it:

$ docker run -it --rm -v /var/run/docker.sock:/var/run/docker.sock sh
/ #

By default, this allows the container to issue any command via that Docker socket (e.g., build and push images to the organization's repositories), which is generally not desirable.

To improve security, the commandList can be configured to restrict the commands that the processes inside the container can issue on the bind-mounted Docker socket. The commandList can be configured as a "deny" list (default) or an "allow" list, depending on your preference.

Each command in the list is specified by its name, as reported by docker --help (e.g., "ps", "build", "pull", "push", etc.) In addition, the following command wildcards are allowed to block an entire group of commands:

For example, the following configuration blocks the build and push commands on the Docker socket:

"commandList": {
  "type": "deny",
  "commands": ["build", "push"]
}

Thus, if inside the container, you issue either of those commands on the bind-mounted Docker socket, they will be blocked:

/ # docker push myimage
Error response from daemon: enhanced container isolation: docker command "/v1.43/images/myimage/push?tag=latest" is blocked; if you wish to allow it, configure the docker socket command list in the Docker Desktop settings or admin-settings.

Similarly:

/ # curl --unix-socket /var/run/docker.sock -XPOST http://localhost/v1.43/images/myimage/push?tag=latest
Error response from daemon: enhanced container isolation: docker command "/v1.43/images/myimage/push?tag=latest" is blocked; if you wish to allow it, configure the docker socket command list in the Docker Desktop settings or admin-settings.

Note that if the commandList had been configured as an "allow" list, then the effect would be the opposite: only the listed commands would have been allowed. Whether to configure the list as an allow or deny list depends on the use case.

Recommendations

• Be restrictive on the list of container images for which you allow bind-mounting of the Docker socket (i.e., the imageList). Generally, only allow this for images that are absolutely needed and that you trust.

• Use the tag wildcard format if possible in the imageList (e.g., <image_name>:*), as this eliminates the need to update the admin-settings.json file due to image tag changes.

• In the commandList, block commands that you don't expect the container to execute. For example, for local testing (e.g., Testcontainers), containers that bind-mount the Docker socket typically create / run / remove containers, volumes, and networks, but don't typically build images or push them into repositories (though some may legitimately do this). What commands to allow or block depends on the use case.

  • Note that all "docker" commands issued by the container via the bind-mounted Docker socket will also execute under enhanced container isolation (i.e., the resulting container uses a the Linux user-namespace, sensitive system calls are vetted, etc.)

Caveats and limitations

• When Docker Desktop is restarted, it's possible that an image that is allowed to mount the Docker socket is unexpectedly blocked from doing so. This can happen when the image digest changes in the remote repository (e.g., a ":latest" image was updated) and the local copy of that image (e.g., from a prior docker pull) no longer matches the digest in the remote repository. In this case, remove the local image and pull it again (e.g., docker rm <image> and docker pull <image>).

• It's not possible to allow Docker socket bind-mounts on local images (i.e., images that are not on a registry) unless they are derived from an allowed image. That's because Docker Desktop pulls the digests for the allowed images from the registry, and then uses that to compare against the local copy of the image.

• The commandList configuration applies to all containers that are allowed to bind-mount the Docker socket. Therefore it can't be configured differently per container.

• The following commands are not yet supported in the commandList:

Note

Docker socket mount permissions do not apply when running "true" Docker-in-Docker (i.e., when running the Docker Engine inside a container). In this case there's no bind-mount of the host's Docker socket into the container, and therefore no risk of the container leveraging the configuration and credentials of the host's Docker Engine to perform malicious activity. Enhanced Container Isolation is capable of running Docker-in-Docker securely, without giving the outer container true root permissions in the Docker Desktop VM.